WO2020099902A1

WO2020099902A1 - Methods and apparatus to perform image analyses in a computing environment

Info

Publication number: WO2020099902A1
Application number: PCT/IB2018/001485
Authority: WO
Inventors: Emilio Almazán; Javier TOVAR VELASCO; Roberto Arroyo; Diego GONZALEZ SERRADOR
Original assignee: The Nielsen Company (Us), Llc
Priority date: 2018-11-13
Filing date: 2018-12-18
Publication date: 2020-05-22
Also published as: US11030769B2; US11715292B2; US20200151902A1; US20210366149A1; WO2020152487A1

Abstract

An apparatus includes that a feature extractor to generate a plurality of image descriptors based on a plurality of retails product tag images corresponding to a retailer category, the plurality of image descriptors including values representative of one or more visual features of the plurality of retail product tag images; a probability density function generator to generate a probability density function of probability values corresponding to the visual features represented in the plurality of image descriptors of the retailer category; a sample selector to select ones of the probability values based on a sample selection algorithm that, when seeded by a retailer seed value corresponding to the retailer category, generates position values identifying positions in the probability density function of the ones of the probability values to he selected; a category signature generator to generate a category signature based on the selected ones of the probability values; and a processor to train a convolutional neural network based on a feature descriptor and at least one of the retail product tag images, the feature descriptor including the category signature concatenated to one of the image descriptors, the training of the convolutional neural network to cause the convolutional neural network to classify the al least one of the retail product tag images as one of a plurality of types of product tags.

Description

METHODS AND APPARATUS TO PERFORM IMAGE ANALYSES IN A COMPUTING ENVIRONMENT

RELATED APPLICATION

|00011 This patent arises from a continuation of PCT Patent

Application No. PCT/ES2018/000078, which was filed on November 13,

2018. PCT Patent Application No. PCT/ES2018/000078 is hereby incorporated herein by reference in its entirety.

FIELD OF THE DISCLOSURE

[0002] This disclosure relates generally to computing systems, and, more particularly, to methods and apparatus to perform image analyses in a computing environment.

BACKGROUND

[0001] Image recognition involves computer-aided techniques to analyze pictures or photographs to determine and/or identify the content of the captured scene (e.g., the recognition of the general subject matter of the scene and/or the recognition of individual objects within the scene). Such techniques are useful in different applications across different industries. For example, retail establishments, product manufacturers, and other business establishments may take advantage of image recognition techniques of photographs of such establishments (e.g., pictures of product shelving) to identify quantities and/or types of products in inventory, to identify shelves that need to be restocked and/or the frequency with which products need restocking, to recognize and read product barcodes, to assess product arrangements and displays, etc.

BRIEF DESCRIPTION OF THE DRAWINGS

[0002] FIG. 1 A is a block diagram of an example apparatus in training mode to train a convolutional neural network (CNN) to classify retail product tags using a category signature.

[0003] FIG. 1 B is an example Gaussian probability distribution function (PDF) generated by a PDF generator of FIG. 1 A.

[0004] FIG. 1C is an example manner of generating a category signature by a category signature generator of FIG. 1 A.

[0005] FIG. 2 is a block diagram of the example apparatus of FIG. 1 A operating in a recognition mode to classify retail product tags using the training CNN and the category signature.

[0006] FIGS. 3A and 3B depict a flowchart representative of machine readable instructions which may be executed to implement the example apparatus of FIGS. 1 A and 2 to operate in the training mode.

[0007] FIGS. 4A and 4B depict flowcharts representative of machine readable instructions which may be executed to implement the example apparatus of FIGS. 1 A and 2 to operate in the recognition mode of FIG. 2.

[0008] FIG. 5 depicts a flowchart representative of machine readable instructions which may be executed to implement the example category signature generator illustrated in FIGS. 1 A and 1C. [0009] FIG. 6 is a block diagram of an example processing platform structured to execute the instructions of FIGS. 3 A, 3B, 4A, 4B, and 5 to implement the example apparatus of FIGS. 1A and 2 to train a CNN and classify retail product tags using the trained CNN and a category signature.

[0010] The figures are not to scale. Wherever possible, the same reference numbers will be used throughout the drawing(s) and accompanying written description to refer to the same or like parts.

DETAILED DESCRIPTION

[0011] In the retail industry, retailers use different types of retail product tags to emphasize or advertise different pricing models for different types of promotions. Across retailers, various different looking retail product tags may be used for the same type of promotion. For example, a first retailer could create a“two for one" retail product tag that is visibly different (e.g., different color, different type font, different logos, different information, different arrangement of information, etc.) from a“two for one” retail product tag of a second retailer.

[0012] When performing retail audits or market analyses for product placements and sales, sorting and categorizing different retail product tags by promotion type and retailer can be performed through human visual inspection using cognitive recognition of persons and human intuition. Such a human- based process is time-consuming and error-prone due to inconsistencies across different human auditors and/or inconsistencies of a same human auditor over time. However, computerizing such a process to recognize different retail product tags of different promotion types and across multiple retailers presents significant challenges.

[0013] Retailer market studies can use deep learning techniques as a way to detect and classify promotions associated with their products. In prior deep learning techniques, a single convolutional neural network (CNN) is trained and utilized specifically for use with a respective retailer to detect or classify retail product tags corresponding to that retailer. However, this is not an easily scalable solution or a solution that is easily updateable or maintained across multiple retailers. For example, a challenge arises when a second retailer or more retailers are introduced into a market study because a second

CNN or more CNNs need to be trained for the second or more retailer categories. As more retailers are added, more CNNs are needed to be trained and maintained. This requires a significant amount of time, supervision, and expertise. In other prior deep learning techniques, a single CNN is trained and utilized for use with two or more respective retailers to detect or classify retail product tags corresponding to the two or more retailers. Challenges arise when there is interference (e.g., similar appearance) between retail product tags of different retailers. For example, a first retailer may have a multi-buy retail product tag image that is of similar appearance to a price reduction retail product tag image of a second retailer. In this example, the CNN fails to properly classify the retailer product tag image of the first retailer due to the interference (e.g., similar appearance) with the retail product tag image of the second retailer.

[0014] As used herein, a categorical entity, a retailer category, or a category refers to a retailer that sells goods and/or services. Categorical entity, retailer category, retailer, and category are used interchangeably throughout.

In examples disclosed herein, a categorical entity, retailer category, and/or category is, but not limited to, a chain of stores owned by a same business entity (e.g., Gap Inc., Banana Republic LLC, Old Navy LLC, Athlete Inc. and

Intermix LLC), a single store (e.g., a small family owned business such as a mom and pop shop), and/or sub-category stores of large retailers such as a big box store, a big box store express, a Best Buy® store, a Best Buy® Mobile store, a Target® store, a Super Target® store, a Walmart® store, a Walmart

Express store, etc.

[0015] Examples disclosed herein overcome the above-mentioned challenges and other challenges by using multi-class classification models in which the final classes (e.g., types of promotions) of retail product tags to be recognized are classified within their own retailer categories. For example, when identifying a type of promotion of a retail product tag of a retailer, examples disclosed herein differentiate between different retailer categories of the retail product tag using respective category signatures generated for corresponding ones of the retailer categories. For example, retailer categories are used to train the CNN so that during a recognition phase, the CNN can differentiate between retailer product tags of different retailer categories. Examples disclosed incorporate the retailer category information into the CNN process by associating the unique category signatures to corresponding retailers based on retailer identifiers or retailer metadata in retailer product tags.

[0016] In examples disclosed herein, a processor trains a CNN based on a feature descriptor and at least one of the retail product tag images. As described in detail below, the feature descriptor includes the category signature concatenated with an image descriptor of the retail product tag image. Training the CNN is to cause the CNN to classify the at least one of the retail product tag images as one of a plurality of types of product tags.

[0017] In examples disclosed herein, a category signature

corresponding to an example retailer category is provided to a CNN to classify input data. In examples disclosed herein, a category signature is a layer of information that is unique for each retailer category. The category signature identifies a category/categorical entity of a product tag. In some examples, the category signature is a value having a particular bit length and could be implemented as a vector, a one-dimensional array of float values (e.g..

floating-point numbers), etc., that is added, augmented, appended, or combined with an image descriptor of a retail product tag to form a feature descriptor that is analyzed and classified by a classifier of the CNN.

[0018] Examples disclosed herein facilitate the aggregating of category information in a CNN while reducing complexity associated with

CNN-based detection and/or classification of retail product tags across multiple retailers. For example, techniques disclosed herein facilitate training a CNN using category signatures corresponding to different categorical entities. By using category signatures for different categorical entities, examples disclosed herein use a single CNN to detect and/or classify retail product tags by promotional type from the different categorical entities. That is, the single CNN is trained to differentiate retail product tags from different retailers based on the category signatures of those retailers.

[0019] examples disclosed herein, a CNN is a network of interconnected neurons (e.g., nodes) arranged or connected to one another in configurations that enable analyzing input data (e.g., image data) to generate different probabilities indicative of likelihoods that different features are present in that input data. In examples disclosed herein, a CNN is provided input image data representative of retail product tags from different categorical entities, and the CNN generates probability values indicative of the likelihoods of the presence of features that correspond to different retail product tags of the categorical entities.

[0020] The category corresponding to a product tag (e g., input data) may be identified by a CNN from metadata located in an image or an image file of a product tag. In examples disclosed herein, metadata is data (e.g., information) that provides information about other data (e.g, a retail product tag). In some examples, metadata of a retail product tag image includes a description, an identifier, camera details such as shutter speed, focal depth, etc., keywords related to the image such as captions, titles, comments, etc., and administrative information such as licensing rights, restrictions on use, etc. In examples disclosed herein, metadata includes a category identifier that identifies a corresponding retailer of a retail product tag. In some examples, the category identifier is geographic location coordinate information of a retail store provided by an image capturing device when capturing images of retail product tags. In such examples, the geographical location information could be used as an identifier of a particular retailer. In other examples, the category identifier could be a value provided to an image capturing device via, for example, user input, scanning a barcode, etc. when capturing images of retail product tags.

[0021] A CNN analyzes input image data to generate probability values indicative of likelihoods that input image data corresponds to different types of retail product tags. In examples disclosed herein, a type of retail product tag is a product tag for a particular type of promotion of a

corresponding product. One type of promotion is a multi-buy promotion (e.g„

“Buy two get one free”,“Get two for the price of one”,“Buy three for $6”, etc.). A second type of promotion is a price reduction promotion (e.g.,“50% off all produce items",“All items are 20% off original price”,“25% OFF”, etc.). Examples disclosed herein may also be employed with other types of promotions.

[0022] FIGS. 1 A and 2 illustrate an example apparatus 100 that employs a CNN 102 to classify retail product tags corresponding to different categorical entities. In FIG. 1 A, the example apparatus 100 is shown in a training mode of a training phase in which category signatures are generated and the CNN 102 is trained to recognize and classify promotional product tags of different retailers using such category signatures. In FIG. 2, the apparatus

100 is shown in a recognition mode of a recognition phase in which the apparatus 100 utilizes category signatures generated during the training mode of FIG. 1 A to analyze input image data of retail product tags and to classify the input image data as representative of types of promotions of different retailers.

[0023] Turning to FIG. 1 A, the example CNN 102 includes a neural network that analyzes input images and learns system responses corresponding to different features in such input images. An input image is a matrix of pixel values that define different colors to collectively represent a visual depiction of a subject. The example CNN 102 has four main operations to analyze the matrix of pixel values. The example operations include a convolution operation, non-linearity operation, a pooling or sub-sampling operation, and a classification operation. The convolution operation extracts features via a filter, from the input image. The filter is a weighted matrix used to analyze the input image matrix. Examples of extracted features include a rounded edge, a circle, a color, etc. The extracted features of the input image are sent to the non-linearity operation of the CNN 102. The non-linearity operation converts the extracted features into a non-linear matrix by replacing negative pixel values with a zero. After the extracted features are non-linearized, the pooling operation samples the extracted features to reduce the dimension of the matrices. The final reduced matrices are then sent to the classification operation of the CNN 102. At the classification operation, the reduced matrices are used to estimate probabilities for the input image. The

probabilities correspond to likely classifications of the input image.

[0024] In some examples, the filters and operations in a CNN 102 are to identify the most important regions of the input image (e.g., the CNN 102 determines the most discriminative features of the input image that make it unique and distinct from the other input images). For example, if the CNN 102 is to classify Coca-Cola® drink label and a Pepsi® drink label. The example

CNN 102 will determine the red color of the Coca-Cola® drink label is a significant feature and therefore the image descriptor of a Coca-Cola® drink label will contain information related to the red color. This is achieved by using filters that get activated in regions with red color. In some examples, there are a plurality of filters, and each filter is to determine different features of the image (e.g., a filter may activate when there is a face, another filter activates when there is a screen, etc.). Eventually, all the information is compacted using a sub-sampling operation, leading to a detailed description of the image (represented by an image descriptor which represents different features in the input image).

[0025] In the example training mode of FIG. 1 A, the CNN 102 is trained based on multiple retail product tag images per retailer category to classify each retail product tag image correctly according to corresponding product promotion type. The example CNN 102 of the apparatus 100 includes an example feature extractor 106 and an example classifier 122. The example apparatus 100 further includes an example retail product tag image database

104, an example image descriptor database 1 10, an example seed generator

1 12, an example category signature generator 1 14, an example feature descriptor generator 1 18, an example retailer ID extractor 126, an example category signature mapper 130, an example category signature database 132, and an example PDF generator 134.

[0026] In the illustrated example, the example apparatus 100 is provided with the feature extractor 106 to obtain retail product tag images from the retail product tag image database 104. The example feature extractor

106 analyzes the images received from the retail product tag image database

104 and extracts particular features of the retail product tag image that make it different relative to a different retail product tag image. For example, the example feature extractor 106 uses the convolution, non-linearity operations, and pooling discussed above. The convolution layer includes a set of filters that convolve over the retail product tag image. The filters are weighted matrices in which the weights in the matrices, originally based on predetermined values, are adjusted in the training mode during backpropagation of the classification operation of the example apparatus 100. The matrices include specified pixel values to extract the particular features of the retail product tag image (e.g., a rounded edge, a color, a character, a logo, etc.) of the target retailer category under analysis. For example, the received retail product tag image may be a [5x5] matrix of pixels arranged in an order that describes the shape, color, size, etc. of the retail product tag image. The example filter may be a [3x3] weighted matrix that convolves (e.g., slides) across the width and the height of the input retail product tag image so that all pixels of the input image are processed at least once (e.g., by one pixel) by the filter. During convolution, the filter generates an output matrix that includes rows and columns of elements (e.g., matrix values) corresponding to pixels of the input image. In the example, the feature extractor 106 computes each element of the output matrix for a corresponding position of the filter during convolution. When the feature extractor 106 applies the example [3x3] filter to the input retail product tag image, the filter processes a [3x3] section of the input retail product tag image pixel values. The pre-determined pixel values of the [3x3] filter and the pixel values of the input image section that is processed is multiplied and the total number added to produce a single element of the processed section. In this manner, the feature extractor 106 generates an output of the final product as a [3x3] matrix of the computed elements of the processed section of the input image. In some examples, the feature extractor

106 includes filters with a plurality of different weighted matrices to extract a plurality of different features of the input image.

[0027] In some examples, the pooling layer processes the input data after the non-linearity operation and convolution layer. The example feature extractor 106 uses the pooling layer to downsample the final computed matrices of the retail product tag image to reduce the spatial size of the image.

For example, if there were ten [3x3] filters applied to the [5x5] target retail product tag image, there would be ten [3x3] output products of the retail product tag image. The feature extractor 106 uses downsampling to reduce the size of the output products of the retail product image for processing, while maintaining the pertinent information regarding the input images in each output product. In some examples, the feature extractor 106 uses max pooling, average pooling, or L2 pooling techniques to downsample.

[0028] The example feature extractor 106 generates the example image descriptor 108 after the pooling operation described above. An image descriptor 108 is a compact representation of the retail product tag image that includes feature values corresponding to the strengths of features detected in the retail product tag image. In the illustrated example, the feature extractor

106 transforms the input image to a one-dimensional array to form the length of the image descriptor 108. As used herein, the length of the image descriptor

108 is a one-dimensional array of size m, where m represents a number of feature values for corresponding features evaluated in an input retail product tag image. Example values m of the image descriptor 108 may be 0.4, 0.2,

0.03, etc. For example, the length of the image descriptor 108 depends on the size of the retail product tag image in pixels and the architecture of the CNN

102. In some examples, the size of the retail product tag image is fixed (e.g.. the retail product tag images are rescaled to a size of 20x20 pixels). As used herein, the architecture of the example CNN 102 is defined as the types of techniques used in three of the four main operations of the CNN 102 (c.g., the convolution operation, the non-linearity operation, and the pooling or sub- sampling operation) to define the final length of the example image descriptor

108. For examples in which a retail product tag image is a size of [20x20] pixels and the architecture of the CNN 102 includes two sub-sampling operations, each sub-sampling operation to reduce the size of the image by half. In this example, the first sub-sampling operation sub-samples and reduces the retail product tag image of [20x20] pixels to [ 1 Ox 10] pixels, and the second sub-sampling operation reduces the retail product tag image of

[ 10x10] pixels to [5x5] pixels. The example image descriptor 108 is flattened to a length of an array of [1x25]. As used herein, flattening a matrix is defined as multiplying the number of rows by the number of columns. In the illustrated example 100, the feature extractor 106 stores the image descriptor 108 in the image descriptor database 1 10, For example, if the example feature extractor

106 receives 1000 images, it will generate 1000 image descriptors and 1000 x m values and store the 1000 image descriptors in the image descriptor database

1 10.

[0029] The example apparatus 100 is provided with the example retailer ID extractor 126 to extract retail product tag metadata from input image data (e.g., retail product tag images) stored in the retail product tag image database 104. The metadata of the retail product tag image identifies the retailer category of the retail product tag image. The example retailer ID extractor 126 stores the extracted metadata as the retailer ID 128 for the retailer category of the retail product tag image in a memory of the example processing platform 600 and provides the retailer ID 128 to the example seed generator 1 12.

[0030] The example image descriptor database 1 10 stores all computed image descriptors in association with corresponding retailer IDs 128 provided by the example retailer ID extractor 126. The example image descriptor database 1 10 notifies the example PDF generator 134 when the example feature extractor 106 computes image descriptors for one or more available retail product tag images of one or more retailer categories.

[0031] The example apparatus 100 is provided with the example PDF generator 134 to generate a probability density function (PDF) based on the image descriptors stored in the image descriptor database 1 10. As used herein, a PDF is a distribution of probabilities of occurrence of different detected visual features across multiple retail product tag images for a particular category. For example, the feature extractor 106 computes 1000 image descriptors and 1000 x m values for 1000 retail product tag images of a first retailer category. The example PDF generator 134 generates a Gaussian PDF

140 that includes probabilities corresponding to multiple features detected in the input images for the retailer category. The example PDF generator 134 models the example Gaussian PDF 140 with a Gaussian distribution to estimate the mean and variance of the magnitudes of the Gaussian PDF 140.

For example, the Gaussian distribution is a bell-shaped curve that is symmetric about the mean of the Gaussian PDF 140 and obtains its maximum value (e.g., the value on the x-axis that is closest to the probability of 1 ). [0032] Referring to FIG. I B, an example Gaussian PDF 140 is shown of a probability distribution generated by the example PDF generator 134. In the example Gaussian PDF 140 the x-axis represents positions (m) of probability values that represent different features in input data (e.g., retail product tag images). The y-axis of the example Gaussian PDF 140 represents probabilities“P(x)^M (from 0 to 1.0) that each feature“x = m" is present in the input data. The example PDF generator 134 provides the example Gaussian

PDF 140 to the example category signature generator 1 14.

[0033] Returning to FIG. 1 A, the example apparatus 100 is provided with the example seed generator 1 12 to generate a retailer seed 136 per category, based on corresponding retailer IDs 128. The retailer seed 136 is for subsequent use by the category signature generator 1 14 to generate a category signature 1 16 for that category as described below in connection with FIG. 1 C.

The example seed generator 1 12 could be implemented by a random-number generator or a pseudo-random number generator to generate unique random seeds for each category so that unique category signatures 1 16 can be generated for each category. For example, the seed generator 1 12 uses the retailer ID 128 to seed the random-number generator or pseudo-random number generator which generates a retailer seed 136 that corresponds to the retailer ID 128 of a corresponding retailer category. In some examples, the seed generator 1 12 may not employ a random or pseudo random number generator and instead generate the retailer seed 136 as a replication of the retailer ID 128. The example seed generator 1 12 stores the unique retailer seed 136 in a memory (c.g., in the image descriptor database 1 10 and/or the category signature database 132). In this manner, the category signature generator 1 14 can use the stored retailer seed 136 at any subsequent time to regenerate the category signature 1 16 for the corresponding category should the category signature 1 16 be lost or destroyed.

100341 The example apparatus 100 is provided with the category signature generator 1 14 to generate category signatures 1 16 based on retailer seeds 136 generated by the seed generator 1 12 and Gaussian PDFs 140 generated by the PDF generator 134. For example, for a retailer category, the category signature generator 1 14 obtains a Gaussian PDF 140 from the PDF generator 134, obtains a retailer seed 136, and uses the retailer seed 136 to generate a category signature 1 16 based on the Gaussian PDF 140.

[0035] Referring to FIG. 1C. the example category signature generator

1 14 is shown generating a category signature 1 16. In the example of FIG. 1C, the category signature generator 1 14 includes an example sample selector 150 and is provided with the Gaussian PDF 140 and the example retailer seed 136 to generate the category signature 1 16.

[0036] The example category signature generator 1 14 of FIG. 1C is provided with the sample selector 150 to randomly select ones of the m probability values from the example Gaussian PDF 140 based on the retailer seed 136 to generate the example category signature 1 16. In some examples, the sample selector 150 selects ones of the m probability values based on a sample selection algorithm that, when seeded by a retailer seed 136 value corresponding to a retailer category, generates position values identifying positions in the Gaussian PDF 140 of the ones of the m probability values to be selected. For example, the sample selector 150 is implemented by a sample- selection algorithm that is seeded by the retailer seed 136 and executed by a processor (e.g., the processor 612 of FIG. 6). The example sample selector 150 uses the retailer seed 136 to initialize the sample-selection algorithm to sample the example Gaussian PDF 140 to obtain a quantity of n samples S_n from ones of the m probability values for use to generate the category signature 1 16.

[0037] In the illustrated example of FIG. 1C, the Gaussian PDF 140 is formed of m probability values corresponding to a quantity of m features in the image descriptors 1 10. The sample selector 150 is seeded with the retailer seed 136 to generate random sample positions corresponding to different ones of the m probability values that are to be used for the n samples of the category signature 116. For example, in FIG. 1C, the first sample (SI) of the category signature 1 16 is selected as the 6^th position probability value (m - 6), and the second sample (S2) is selected as the 8^,h probability value (iw^~8) of the

Gaussian PDF 140.

[0038] The example sample selector 150 is deterministic in that when seeded with the same value for the retailor seed 136, the sample selector 150 will deterministically output the same values for corresponding positions of the m probability values of the Gaussian PDF 140 for use in forming the category signature 1 16. As such, the category signature 1 16 can be consistently reproduced based on the same retailer seed 136 (e.g., if the category signature 1 16 is lost or deleted).

[0039] In some examples, the sample selector 150 draws the n samples from the Gaussian PDF 140 (e.g., selects ones of the positions of the probability values of the Gaussian PDF 140) from a range or group of features m having the highest probabilities of occurrence. For example, Si 144, S2 146, and S_n 148 are all example samples corresponding to features m having highest probabilities of occurring in the retail product tag of a retailer category. In this example, n represents a number of samples equal to the length of the category signature 1 16. For example, the category signature 1 16 may be a [ 1x25] matrix of samples, in which the“25” in the matrix represents the length or quantity n of samples.

[0040] Referring again to FIG. I A, since the input data is a group of image descriptors 108 from the image descriptor database 1 10 for retail product tag images of the same retailer category, the Gaussian PDF 140 of the image descriptors 108 will be in the same numeric range as the feature values of each individual image descriptor 108 for the same retailer category. In addition, the resulting category signature 1 16 based on the Gaussian PDF 140 is in the same numeric range as the image descriptors 108 of retail product tags for the same retail category.

[0041] In the illustrated example, the example category signature generator 1 14 provides the category signature 1 16 to the feature descriptor generator 1 18. In the illustrated example, the feature descriptor generator 1 18 obtains the image descriptor 108 of the same retailer category from the example feature extractor 106 and/or the example image descriptor database

1 10. The example feature descriptor generator 1 18 generates a feature descriptor 120 based on the image descriptor 108 and the category signature

1 16. For example, the feature descriptor generator 1 18 concatenates the matrix

(e.g., a [1x25] matrix) that forms the image descriptor 108 and the matrix

(e.g., a [1x25] matrix) that forms the category signature 1 16. In this manner, the feature descriptor 120 is twice the length (e.g., a [1x50] matrix) of the image descriptor 108.

[0042] As described above, the numeric range of values in the category signature 116 is the same range as values in the image descriptor 108.

Keeping the values in the category signature 1 16 and the values in the image descriptor 108 in the same numeric range is useful to facilitate the classifier

122 to properly process the resulting feature descriptor 120. That is, values for the example category signature 1 16 that are not commensurate in numeric range with the values of the image descriptor 108 decreases the accuracy of the CNN 102. For example, the recognition sensitivity of the CNN 102 to features in the numeric range of values in the image descriptor 108 is decreased if the values of the category signature 1 16 are outside that numeric range because the filters of the feature extractor 106 would need to adapt to process the larger of numeric values of the category signature 1 16, thereby decreasing the sensitivity of the CNN 102 to the features of a retail product tag image represented by values of a smaller numeric range. [0043] The apparatus 100 is provided with the example classifier 122 to classify the provided feature descriptor 120 (e,g., determines output probabilities of feature descriptors 120 for different classes). During training, the example classifier 122 generates multiple outputs over multiple iterations until the feature recognition of the classifier 122 improves such that the output of the classifier 122 is acceptable in correctly classifying input images of retail product tags. For example, the classifier 122 illustrated in the CNN 102 may initially output unacceptable probabilities for two classes (e.g., promotional product tag types such as a multi-buy promotional product tag, a price reduction promotional product tag, etc.) of the feature descriptor 120 at the beginning of a training phase. After multiple training iterations and

backpropagation (e.g., corrective feedback) based on the one or more training retail product tag images, the example classifier 122 outputs probabilities of promotional product tag types that are acceptable. An algorithm is used to calculate the total error at the output of the classifier 122 (e.g., a summation over the two classes). Backpropagation is then used to calculate gradients of the error with respect to all weights (e.g., the filters used during the convolution operation) in the CNN in training mode 102. After

backpropagation of the output probabilities, the weights in the filters of the feature extractor 106 are adjusted (e.g., the [3x3] matrix of pixel values that represent a filter of the feature extractor 106 may be rearranged, adjusted, removed, etc.) to help better predict the class of the feature descriptor 120. [0044] The apparatus 100 is provided with the example category signature mapper 130 to map the retailer ID 128 of a retailer category with the category signature 116 for that retailer category. For example, the category signature mapper 130 stores the category signature 1 16 in association with the retailer ID 128 (e.g., in a same mapping record in the example category signature database 132) to establish a correspondence between the two items of information.

|004S| The apparatus 100 is provided with the example category signature database 132 to store the category signatures that are mapped with their corresponding retailer IDs. For example, the category signature mapper

130 may provide the mapped category signature 1 16 and retailer ID 128 to the category signature database 132 for storing in a corresponding mapping record.

[0046] FIG. 2 is a block diagram of the example apparatus 100 of FIG.

1 A operating in a recognition mode. The example apparatus 100 of FIG. 2 may be implemented using the CNN 102 after the training mode of FIG. 1 A.

In the recognition mode of FIG. 2, the CNN 102 classifies a retail product tag.

[0047] In some examples, during the recognition mode, the CNN 102 is used to classify retail product tags in real time as images of these product tags are captured by a device (e.g., a mobile phone, a tablet device, a digital still camera, etc ). In other examples, the CNN 102 is used to classify retail product tags during a post process during some later time after multiple images of multiple retail product tags have been captured and stored. In any case, the CNN 102 may be implemented by the same device that captured the images or by a separate device that obtains the images from the capture device. In the example recognition mode of FIG. 2, the CNN 100 has been trained to classify input retail product tag images from a plurality of retailer categories and output probability values indicative of the input retail product tag images corresponding to different types of retail product tags. The example apparatus 100 includes the example feature extractor 106, the example retailer

ID extractor 126, the example category signature database 132, the example feature descriptor generator 1 18, and the example classifier 122. In the illustrated example, the apparatus 100 generates an example first output probability value 210 from zero to one (e.g., one being the highest likelihood) that an input retail product tag corresponds to a first type of promotional product tag. The example apparatus 100 also generates an example output probability 212 value that represents a likelihood from zero to one (e.g., one being the highest likelihood) that an input retail product tag corresponds to a second type of retail product tag.

[0048] The apparatus 100 of FIG. 2 is provided with the example retailer categories 202. The retailer categories 202 may be identified in images based on retailer category metadata (e.g., data and/or information that can be used to identify a retailer category). The example retailer categories 202 are identified by their fill type (e.g., slanted lines, checkered lines, diamond shaped lines, etc.). For example, the slanted lines of retailer category 1 may represent the metadata of a large corporate department store, the checkered lines of retailer category 2 may represent the metadata of a mom and pop shop, the diamond shaped lines may represent the metadata of a chain of stores owned by a same business entity, etc.

[0049] The example apparatus 100 in the recognition mode of FIG. 2 begins when the example apparatus 100 obtains a retail product tag image 204 from the retail product tag image database 104. In the recognition mode of

FIG. 2, the apparatus 100 is to determine whether the retail product tag image

204 corresponds to a first type of retail product tag or a second type of retail product tag. In the illustrated example of FIG. 2, the apparatus 100 processes a retail product tag image from category I . However, the same techniques disclosed may be used to classify product tag images from any category.

[0050] The apparatus 100 is provided with the example retailer ID extractor 126 to retrieve the retailer ID (e.g., the retailer ID 128 of FIG. 1 A) from metadata of the retail product tag image 204. The example apparatus 100 is also provided with the category signature database 132 to provide the category signatures for the retailer categories (e.g., the category signature 1 16 generated in the apparatus 100 in training mode of FIG. 1 A). For example, the retailer ID extractor 126 may provide the retailer ID 128 corresponding to the retail product tag image 204 to the category signature database 132. The example category signature database 132 may use the retailer ID 128 to scan a lookup table in the category signature database 132 that has mapped a category signature with the retailer ID 128. [0051] The apparatus 100 is provided with the example feature extractor 106 to generate an image descriptor 216 representative of the retail product tag image 204. The image descriptor 216 includes feature values corresponding to the strengths of features detected in the retail product tag image 204. The image descriptor 216 is similar in structure to the image descriptor 108 of FIG. 1 A. In some examples, the feature extractor 106 analyzes the retail product tag image 204 and extracts particular features of the image 204 that, for example, differ from features of a different retail category product tag image of the same or different retailer category. Example operations of the feature extractor 106 to generate the example image descriptor 216 are described above in connection with FIG. 1 A. The feature extractor 106 of FIG. 2 utilizes filters that are weighted to allow the example classifier 122 to generate probability values indicative of likelihoods of different types of retail product tags. For example, the feature extractor 106 may generate an image descriptor 216 using filters of the same size (e.g., a [3x3] filter) but with different pixel values corresponding to the retail product tag image 204. As described in the above examples and to facilitate further description of the disclosed examples, the example image descriptor 216 of

FIG. 2 is referred to as a one-dimensional array of size [1 x10].

[0052] The apparatus 100 is provided with the example feature descriptor generator 1 18 to generate a feature descriptor 218 corresponding to the retail product tag image 204 of a retailer category by concatenating the image descriptor 216 and the category signature 1 16. For example, the feature descriptor generator 1 18 obtains the category signature 1 16 of the retailer corresponding to the retail product tag image 204 from the category signature database 132 and obtains the image descriptor 216 of the retail product tag image 204 from the feature extractor 106. In the illustrated example, the example category signature database 132 provides the category signature 1 16 to the example feature descriptor generator 1 18 when it has determined the category signature 1 16 corresponding to the retailer ID 128 from the retailer

ID extractor 126 that was mapped and stored by the category signature mapper

130 of the apparatus 100 of FIG. 1 A. In the illustrated example, the feature descriptor 218 is similar in structure to the feature descriptor 120 of FIG. 1 A.

[0053] In the illustrated example, the feature descriptor generator 1 18 generates the feature descriptor 218 to be twice the bit length of the image descriptor 216. For example, the image descriptor 216 is an array of size

[1x25] and the example category signature 1 16 associated with the category of the image descriptor 108 is also an array of size [1x25]. The example feature descriptor generator 118 concatenates the two arrays corresponding to the image descriptor 216 and the category signature 1 16 to generate the feature descriptor 218. The resulting feature descriptor 218 is a size [ 1 x50],

[0054] The example feature descriptor 218 includes data to facilitate properly classifying the retail product tag image 204 by the classifier 122 into a correct class type of retail product tag. The example classifier 122 receives the feature descriptor 218 and outputs a probability value indicative of a class corresponding to the retail product tag image 204. [0055] The example classifier 122 is a non-linear model. The example classifier 122 uses a logistic regression with cost functions to generate the non-linear model. That is, the classifier 122 uses such logistic regression with cost functions to classify the example feature descriptor 218. Logistic regression is a classification algorithm used to assign observations to a discrete set of classes. Logistic regression transforms its output using the logistic sigmoid fbnction to return a probability value which can then be mapped to two or more discrete classes. The logistic sigmoid function is a mathematical function that has an input of a real value and outputs another value between 0 and 1, and the output function is non-linear. In some examples, the classifier 122 utilizes this function because the output probability needs to be between 0 and 1. In the illustrated example of FIG. 2, the output probability value 210 that the retail product tag image 204 is a first type of retail product tag is 0.81 and the probability value 212 that the retail product tag image 204 is a second type of retail product tag is 0.19. Therefore, it is more likely that the class of the retail product tag image 204 is a first type of retail product tag, because the output probability of 0.81 is closer to 1 than the output probability of 0.19 of the example second type of retail product tag.

[0056] The apparatus 100 is provided with the example report generator 222 to generate a report based on the classification results provided by the output probability values 210 and 212. For example, the report generator 222 displays the report to a user, uses the report to determine the price of a product requested by a customer, updates the product prices in a retailer sore and/or website, etc. In some examples, the report generator 222 may be a separate entity of the CNN 102, a computer process, etc.

[0057] While an example manner of implementing the apparatus 100 of FIG. 1 A is illustrated in FIGS. I A, 1 C, and 2, one or more of the elements, processes and/or devices illustrated in FIGS. 1 A and 2 may be combined. divided, re-arranged, omitted, eliminated and/or implemented in any other way. Further, the example CNN 102, the example retail product tag image database 104, the example feature extractor 106, the example image descriptor database 1 10, the example seed generator 1 12, the example PDF generator

134, the example category signature generator 114, the example feature descriptor generator 1 18, the example classifier 122, the example retailer ID extractor 126, the example category signature mapper 130, the example category signature database 132, the example sample selector 150, and/or, more generally, the example apparatus 100 of FIGS. 1 A and 2 may be implemented by hardware, software, firmware and/or any combination of hardware, software and/or firmware. Thus, for example, any of the example

CNN 102, the example retail product tag image database 104, the example feature extractor 106, the example image descriptor database 1 10, the example seed generator 112, the example PDF generator 134, the example category signature generator 1 14, the example feature descriptor generator 1 18, the example classifier 122, the example retailer ID extractor 126, the example category signature mapper 130, the example category signature database 132, the example sample selector 150, and/or, more generally, the example apparatus 100 could be implemented by one or more analog or digital circuits), logic circuits, programmable processors), programmable control!er(s), graphics processing unit(s) (GPU(s)), tensor processing unit(s)

(TPU(s)), digital signal processors) (DSP(s)), application specific integrated circuit(s) (ASIC(s)), programmable logic device(s) (PLD(s)) and/or field programmable logic device(s) (FPLD(s)). When reading any of the apparatus or system claims of this patent to cover a purely software and/or firmware implementation, at least one of the example CNN 102, the example retail product tag image database 104, the example feature extractor 106, the example image descriptor database 1 10, the example seed generator 1 12, the example PDF generator 134, the example category signature generator 1 14, the example feature descriptor generator 1 18, the example classifier 122, the example retailer ID extractor 126, the example category signature mapper 130, the example category signature database 132, and/or the example sample selector 150, is/are hereby expressly defined to include a non-transitory computer readable storage device or storage disk such as a memory, a digital versatile disk (DVD), a compact disk (CD), a Blu-ray disk, etc. including the software and/or firmware. Further still, the example apparatus 100 of FIGS.

1 A and 2 may include one or more elements, processes and/or devices in addition to, or instead of, those illustrated in FIGS. 1 A and 2, and/or may include more than one of any or all of the illustrated elements, processes and devices. As used herein, the phrase“in communication,” including variations thereof, encompasses direct communication and/or indirect communication through one or more intermediary components, and does not require direct physical (e.g., wired) communication and/or constant communication, but rather additionally includes selective communication at periodic intervals, scheduled intervals, aperiodic intervals, and/or one-time events.

[0058] Flowcharts representative of example machine readable instructions for implementing the example apparatus 100 in the training mode of FIG. 1 A are shown in FIGS. 3 A, 3B, and 5 and a flowchart representative of example machine readable instructions for implementing the example apparatus 100 operating in the recognition mode of FIG. 2 is shown in FIGS.

4A and 4B. The machine readable instructions may be one or more executable program(s) or portion(s) of one or more executable program(s) for execution by a computer processor such as the processor 612 shown in the example processor platform 600 discussed below in connection with FIG. 6. The program(s) may be embodied in software stored on a non-transitory computer readable storage medium such as a CD-ROM, a floppy disk, a hard drive, a

DVD, a Blu-ray disk, or a memory associated with the processor 612, but the entirety of the program(s) and/or parts thereof could alternatively be executed by a device other than the processor 612 and/or embodied in firmware or dedicated hardware. Further, although the example program(s) i s/are described with reference to the flowcharts illustrated in FIGS. 3 A, 3B, 4A, 4B, and 5, many other methods of implementing the example apparatus 100 in the training mode of FIG. 1 A and the example apparatus 100 in the recognition mode of FIG. 2 may alternatively be used. For example, the order of execution of the blocks may be changed, and/or some of the blocks described may be changed, eliminated, or combined. Additionally, or alternatively, any or all of the blocks may be implemented by one or more hardware circuits

(c.g., discrete and/or integrated analog and/or digital circuitry, an FPGA, an

ASIC, a comparator, an operational-amplifier (op-amp), a logic circuit, etc.) structured to perform the corresponding operation without executing software or firmware.

[0059] As mentioned above, the example processes of FIGS. 3A, 3B,

4A, 4B, and 5 may be implemented using executable instructions (e.g., computer and/or machine readable instructions) stored on a non-transitory computer and/or machine readable medium such as a hard disk drive, a flash memory, a read-only memory, a compact disk, a digital versatile disk, a cache, a random-access memory and/or any other storage device or storage disk in which information is stored for any duration (e.g., for extended time periods, permanently, for brief instances, for temporarily buffering, and/or for caching of the information). As used herein, the term non-transitory computer readable medium is expressly defined to include any type of computer readable storage device and/or storage disk and to exclude propagating signals and to exclude transmission media.

[0060] The training program of FIGS. 3A and 3B begins at block 302

(FIG. 3 A) at which the example apparatus 100 selects a retailer category to train the CNN 102 (FIG. 1 A) to classify retail product tag images of the selected retailer category. The example apparatus 100 obtains a retail product tag image of the selected retailer category (block 304). For example, the retailer ID extractor 126 and the feature extractor 106 of FIG. 1A obtain the retail product tag image from the example retail product tag image database

104 (FIG. 1 A). The example retailer ID extractor 126 retrieves metadata firom the retail product tag image (block 306). For example, the extracted metadata is the retailer ID 128 (FIG. 1 A) indicative of the retailer category.

[0061] The example feature extractor 106 then determines the features of the retail product tag image (block 308). In some examples, the feature extractor 106 utilizes a convolution operation plus a max pooling operation to determine the features of the product tag image. In the illustrated example, the operation of block 308 is executed once, twice, or a plurality of times for any number of features that may be present in the input image. The feature extractor 106 generates an image descriptor 108 (FIG. 1 A) based on the features (block 310).

[0062] The example feature extractor 106 stores the example image descriptor 108 in the image descriptor database 1 10 (block 312). The example apparatus 100 determines if there is another retail product tag image of the selected retailer category to be analyzed (block 314). For example, the feature extractor 106 or the retailer ID extractor 126 may query the retail product tag image database 104 to determine whether any retail product tag images remain to be processed. If there is another retail product tag image to process, control returns to block 308. In this manner, multiple iterations of blocks 308, 310,

312, 314 can be performed until all or multiple (e.g., a sufficient quantity for satisfactory training of the CNN 102) of the retail product tag images in the retail product tag image database 104 for the selected retailer category have been processed for training.

[0063] Turning to FIG. 3B, the example PDF generator 134 generates a Gaussian PDF 140 (block 316) based on the image descriptors 108. For example, the PDF generator 134 is provided with a plurality of image descriptors 108 from the image descriptor database 1 10 and generates a

Gaussian distribution 140 of probabilities of occurrence of different detected visual features of the corresponding image descriptors 108 for the selected retailer category. The example seed generator 1 12 generates a retailer seed 136

(block 318) per retailer category based on the retailer ID 128. For example, the seed generator 1 12 is provided with the retailer ID 128 and generates a retailer seed 136 that corresponds to the retailer ID 128.

[0064] The example category signature generator 1 14 (FIG. 1 A) generates a category signature 1 16 (block 320). For example, the category signature generator 1 14 selects n samples from the Gaussian PDF 140 as described above in connection with FIG. 1C. An example process that may be used to implement block 320 is described below in connection with FIG. 5.

100651 The example category signature mapper 130 maps the category signature 1 16 to the retailer ID 128 (block 322) of the retailer category. For example, the category signature mapper 130 obtains the generated category signature 1 16 from the category signature generator 1 14 and obtains the example retailer ID 128 of the selected retailer category from the example retailer ID extractor 126. The example category signature mapper 130 stores the mapped category signature 1 16 and the retailer ID 128 (block 324) in the example category signature database 132. For example, the category signature mapper 130 stores the category signature 1 16 in association with the retailer

ID 128 in a mapping record in the category signature database 132.

[0066] The example feature descriptor generator 1 18 (FIG. 1 A) generates a feature descriptor 120 (block 326). For example, the example feature descriptor 120 (FIG. 1A) concatenates the example category signature

1 16 and the example image descriptor 108 to forni the feature descriptor 120 as an array twice the size of the example image descriptor 108.

[0067] The example classifier 122 (FIG. 1 A) of the CNN 102 is trained based on the feature descriptor 120 (block 328). For example, the processor 612 (FIG. 6) may train the classifier 122 to output correct probabilities via backpropagation. Backpropagation is used to calculate gradients of an error with respect to all weights (e.g., the filters used during the convolution operation) in the apparatus 100 in training mode. After backpropagation of the output probabilities, the weights in the filters of the classifier 122 are adjusted until the errors corresponding to the weights are reduced to an acceptable minimum error value (e.g., below a threshold). The examples process of FIGS. 3 A and 3B then ends.

[0068] In some examples, the processes of FIGS. 3A and 3B may be repeated a plurality of times for the same retailer category and/or different retailer categories. The processes may be repeated a plurality of times until the CNN 102 in training mode has been trained to output acceptable probabilities

(e.g., based on acceptable validation results) for each class of each retailer category.

[0069] The recognition program of FIGS. 4A and 4B begins at block

404 (FIG. 4A) at which the example apparatus 100 obtains a retail product tag image 204 (block 404). For example, the retailer ID extractor 126 and the feature extractor 106 of FIG. 2 obtain the retail product tag image 204 from the example retail product tag image database 104 (FIG. 2).

]0070] The example retailer ID extractor 126 obtains a retailer ID

(block 406). For example, the retailer ID extractor 126 retrieves a retailer ID as metadata from the retail product tag image. The retail ID is indicative of a retailer category for which to obtain a category signature 1 16 from the example category signature database 132. The example feature descriptor generator 1 18 obtains the category signature 1 16 (block 408). For example, the feature descriptor generator 1 18 obtains the category signature 1 16 from the category signature database 132 based on the retailer ID.

[0071] The example feature extractor 106 determines the features of the retail product tag image 204 (block 410). In some examples, the feature extractor 106 utilizes the convolution operation plus the pooling operation to compute the features from the retail product tag image 204. In some examples. this operation is executed once, twice, or a plurality of times for one or more features that may be present in the retail product tag image 204. The example feature extractor 106 generates an image descriptor 216 based on the features

(block 412).

[0072] The example feature descriptor generator 1 18 (FIG. 2) generates a feature descriptor 218 (block 414) based on the category signature

1 16 and the image descriptor 216. For example, the feature descriptor generator 1 18 generates the example feature descriptor 218 (FIG. 2) by concatenating the example category signature 1 16 with the example image descriptor 216. The example feature descriptor 218 (FIG. 2) is an array twice the size of the example image descriptor 216.

[0073] The example feature descriptor generator 1 18 stores the feature descriptor 218 in a memory (block 416). For example, the feature descriptor generator 1 18 provides the feature descriptor 218 to a memory accessible by the classifier 122 (FIG. 2). In some examples, the memory is the implemented by one or more of the local memory 613, the volatile memory 614, the nonvolatile memory 616, or the mass storage 628 of FIG. 5.

[0074] The example classifier 122 generates a first probability for a first type of promotional product tag (block 418). The example classifier generates a second probability for a second type of promotional product tag

(block 420). For example, the example classifier 122 outputs probabilities for two classes, which include a multi-buy retail product tag and a price reduction retail product tag. For example, the classifier 122 may utilize a logistic regression function as described above to determine the probability values. [0075] Based on the probability values from the logistic regression function, the example classifier 122 determines the type of promotional price tag of the retail product tag image (block 422). For example, the example classifier 122 outputs a probability value 210 (FIG. 2) of 0.81 for a first type of promotional product tag and a probability value 212 (FIG. 2) of 0.19 for a second type of promotional product tag. Based on such example probabilities, it is most likely that the class of the retail product tag image 204 is a first type of promotional product tag, because the output probability of 0.81 of the example first type of promotional product tag is closer to the value of 1 than the output probability value of 0.19 of the example second type of promotional product tag. The example classifier 122 stores the output results (block 424).

For example, the classifier 122 store the probabilities in memory in association with the retail product tag image and/or in association with an identifier (e.g., a file name) of the retail product tag image. The apparatus 100 determines if there is another retail product tag image to analyze (block 426). For example, the apparatus 100 queries the retail product tag image database 104 for any other retail product tag image to be analyzed.

100761 If example apparatus 100 determines that another product tag image is to be analyzed, control returns to block 404 in FIG. 4A, In this manner, multiple iterations of blocks 404, 406, 408, 410, 412, 414, 416, 418,

420, 422, and 424 can be performed. If another product tag image is not available to be analyzed, the example report generator 222 (FIG. 2) generates a report based on the classification results (block 428) of the analyzed retail product tag images. For example, report generator 222 displays the report to a user, uses the report to determine the price of a product requested by a customer, updates the product prices in a retailer sore and/or website, etc.

[0077] Turning now to FIG. 5, an example flowchart represents a category signature generation program that may be used to implement the example category signature generator 1 14 of FIGS. 1 A and 1C. The process of

FIG. 5 begins at block 502 at which the example category signature generator

1 14 receives the retailer seed 136 and the Gaussian PDF 140 corresponding to the selected category. The example category signature generator 1 14 initializes the example sample selector 150 with the example retailer seed 136 (block

504) to randomly select samples from the example Gaussian PDF 140. For example, the sample selector 150 is seeded with the retailer seed 136 to determine positions of ones of the m values of the Gaussian PDF 140 to select as the n samples for forming the retailer category signature 1 16 as described above in connection with FIG. 1C.

100781 The example sample selector 150 selects n samples from the example Gaussian PDF 140 (block 506). For example, the sample selector 150 selects ones of m positions of the probability values of the Gaussian PDF 140 based on n sample position values generated by the sample selector 150. The example category signature generator 1 14 generates a one-dimensional array with the n samples from the example Gaussian PDF 140 (block 508). For example, the category signature generator 1 14 retrieves the n samples selected by the example sample selector 150 at block 506 and combines them into a one-dimensional array of a quantity of n elements.

100791 The example category signature generator 1 14 stores the onedimensional array as a category signature 1 16 (block 510) for future use by the example CNN 102 (FIGS. 1 A and 2). For example, the category signature generator 1 14 stores the category signature 1 16 in the category signature database 132. In this manner, the feature descriptor generator 1 18 can access the category signature 1 16 from the category signature generator 1 14 or from the category signature database 132 to generate a feature descriptor (e.g., the feature descriptor 218 of FIG. 2). The example process of FIG. 5 ends, and control return to an example calling function or process such as the example process of FIG. 3B.

[0080] As mentioned above, the example processes of FIGS. 3A, 3B,

4A, 4B, and 5 may be implemented using executable instructions (e.g., computer and/or machine readable instructions) stored on a non-transitory computer and/or machine readable medium such as a hard disk drive, a flash memory, a read-only memory, a compact disk, a digital versatile disk, a cache, a random-access memory and/or any other storage device or storage disk in which information is stored for any duration (e.g._s for extended time periods, permanently, for brief instances, for temporarily buffering, and/or for caching of the information). As used herein, the term non-transitory computer readable medium is expressly defined to include any type of computer readable storage device and/or storage disk and to exclude propagating signals and to exclude transmission media.

[0081]“Including” and“comprising” (and all forms and tenses thereof) are used herein to be open ended terms. Thus, whenever a claim employs any form of“include” or“comprise” (e.g., comprises, includes. comprising, including, having, etc.) as a preamble or within a claim recitation of any kind, it is to be understood that additional elements, terms, etc. may be present without falling outside the scope of the corresponding claim or recitation. As used herein, when the phrase "at least” is used as the transition term in, for example, a preamble of a claim, it is open-ended in the same manner as the term "comprising" and“including” are open ended. The term

“and/or” when used, for example, in a form such as A, B, and/or C refers to any combination or subset of A, B, C such as (1) A alone, (2) B alone, (3) C alone, (4) A with B, (5) A with C, (6) B with C, and (7) A with B and with C.

As used herein in the context of describing structures, components, items, objects and/or things, the phrase "at least one of A and B" is intended to refer to implementations including any of (1) at least one A, (2) at least one B, and

(3) at least one A and at least one B. Similarly, as used herein in the context of describing structures, components, items, objects and/or things, the phrase

"at least one of A or B” is intended to refer to implementations including any of (1) at least one A, (2) at least one B, and (3) at least one A and at least one

B. As used herein in the context of describing the performance or execution of processes, instructions, actions, activities and/or steps, the phrase "at least one of A and B" is intended to refer to implementations including any of (1 ) at least one A, (2) at least one B, and (3) at least one A and at least one

B, Similarly, as used herein in the context of describing the performance or execution of processes, instructions, actions, activities and/or steps, the phrase

"at least one of A or B" is intended to refer to implementations including any of (1 ) at least one A, (2) at least one B, and (3) at least one A and at least one

B.

[0082] FIG. 6 is a block diagram of an example processor platform

600 structured to execute the instructions of FIGS 3 A, 3B, 4A, 4B, and 5 to implement the apparatus 100 of FIGS. 1 A, 1C, and 2. The processor platform

600 can be, for example, a server, a personal computer, a workstation, a self- learning machine (e.g., for neural network processing), a mobile device (e.g., a cell phone, a smart phone, a tablet such as an iPad™), a personal digital assistant (PDA), an Internet appliance, a headset or other wearable device, or any other type of computing device.

[0083] The processor platform 600 of the illustrated example includes a processor 612. The processor 612 of the illustrated example is hardware.

For example, the processor 612 can be implemented by one or more integrated circuits, logic circuits, microprocessors, GPUs, DSPs, or controllers from any desired family or manufacturer. The hardware processor may be a

semiconductor based (e.g., silicon based) device. In this example, the processor 612 implements the example feature extractor 106, the example retailer ID extractor 126, the example seed generator 1 12, the example category signature generator 1 14, the example category signature mapper 130, the example feature descriptor generator 1 18, the example classifier 122, the example report generator 222, the example PDF generator 134, and the example sample selector 150.

100841 The processor 612 of the illustrated example includes a local memory 613 (c.g., a cache). The processor 612 of the illustrated example is in communication with a main memory including a volatile memory 614 and a non-volatile memory 616 via a bus 618. The volatile memory 614 may be implemented by Synchronous Dynamic Random Access Memory (SDRAM),

Dynamic Random Access Memory (DRAM), RAMBUS® Dynamic Random

Access Memory (RDRAM®) and/or any other type of random access memory device. The non-volatile memory 616 may be implemented by flash memory and/or any other desired type of memory device. Access to the main memory

614, 616 is controlled by a memory controller.

100851 The processor platform 600 of the illustrated example also includes an interface circuit 620. The interface circuit 620 may be

implemented by any type of interface standard, such as an Ethernet interface, a universal serial bus (USB), a Bluetooth® interface, a near field

communication (NFC) interface, and/or a PCI express interface.

[0086] In the illustrated example, one or more input devices 622 are connected to the interface circuit 620. The input device(s) 622 permit(s) a user to enter data and/or commands into the processor 612. The input device(s) can be implemented by, for example, an audio sensor, a microphone, a camera (still or video), a keyboard, a button, a mouse, a touchscreen, a trackpad, a trackball, isopoint and/or a voice recognition system.

[0087] One or more output devices 624 are also connected to the interface circuit 620 of the illustrated example. The output devices 624 can be implemented, for example, by display devices (e.g., a light emitting diode

(LED), an organic light emitting diode (OLED), a liquid crystal display

(LCD), a cathode ray tube display (CRT), an in-place switching (IPS) display, a touchscreen, etc.), a tactile output device, a printer and/or speaker. The interface circuit 620 of the illustrated example, thus, typically includes a graphics driver card, a graphics driver chip and/or a graphics driver processor.

10088] The interface circuit 620 of the illustrated example also includes a communication device such as a transmitter, a receiver, a transceiver, a modem, a residential gateway, a wireless access point, and/or a network interface to facilitate exchange of data with external machines (e.g., computing devices of any kind) via a network 626. The communication can be via, for example, an Ethernet connection, a digital subscriber line (DSL) connection, a telephone line connection, a coaxial cable system, a satellite system, a line-of-site wireless system, a cellular telephone system, etc.

[0089] The processor platform 600 of the illustrated example also includes one or more mass storage devices 628 for storing software and/or data. Examples of such mass storage devices 628 include floppy disk drives, hard drive disks, compact disk drives, Blu-ray disk drives, redundant array of independent disks (RAID) systems, and digital versatile disk (DVD) drives. [0090] The machine executable instructions 632 of FIGS 3A-4B may be stored in the mass storage device 628, in the volatile memory 614, in the non-volatile memory 616, and/or on a removable non-transitory computer readable storage medium such as a CD or DVD.

[0091] From the foregoing, it will be appreciated that example methods, apparatus and articles of manufacture have been disclosed that improve the performance of a CNN by generating and using a category signature as an extra layer of information to help the CNN classify an inputted retail product tag image without needing multiple CNNs for different retailer categories and without needing user-input to identify a retailer category of an input retail product tag image during a recognition phase. In this manner, the disclosed methods, apparatus and articles of manufacture improve the efficiency of using a computing device by significantly reducing the number of CNNs and user-input needed to classify retail product tag images across a plurality of retailers. The disclosed methods, apparatus and articles of manufacture are accordingly directed to one or more improvement(s) in the functioning of a computer.

[0092] Example 1 is an apparatus including a feature extractor to generate a plurality of image descriptors based on a plurality of retail product tag images corresponding to a retailer category, the plurality of image descriptors including values representative of one or more visual features of the plurality of retail product tag images, a probability density function generator to generate a probability density function of probability values corresponding to the visual features represented in the plurality of image descriptors of the retailer category, a sample selector to select ones of the probability values based on a sample selection algorithm that, when seeded by a retailer seed value corresponding to the retailer category, generates positron values identifying positions in the probability density function of the ones of the probability values to be selected, a category signature generator to generate a category signature based on the selected ones of the probability values, and a processor to train a convolutional neural network based on a feature descriptor and at least one of the retail product tag images, the feature descriptor including the category signature concatenated to one of the image descriptors, the training of the convolutional neural network to cause the convolutional neural network to classify the at least one of the retail product tag images as one of a plurality of types of product tags.

[0093] 1° Example 2, the subject matter of Example 1 can optionally include that probability density function generator is to model the probability density function of the probability values with a Gaussian distribution, the

Gaussian distribution to approximate a real distribution.

[0094] In Example 3, the subject matter of any one of Examples 1 -2 can optionally include a seed generator to generate the retailer seed value based on a retailer identifier of the retailer category.

[0095] In Example 4, the subject matter of any one of Examples 1-3 can optionally include that the seed generator is implemented by a random- number generator to generate the retailer seed as a unique random seed. [0096] In Example 5, the subject matter of any one of Examples 1 -4 can optionally include the category signature is a length representing a number of samples selected by the sample selector, the samples corresponding to ones of the probability values.

[0097] In Example 6, the subject matter of any one of Examples 1-5 can optionally include that the category signature is reproduceab!e based on the same retailer seed value.

[0098] In Example 7, the subject matter of any one of Examples 1-6 can optionally include a memory to store the category signature in association with a retailer identifier corresponding to the retailer category, the category signature different from a second retailer category generated by the category signature generator for a second retailer.

[0099] Example 8 is a non-transitory computer readable storage medium comprising instructions that, when executed, cause a processor to at least: generate a plurality of image descriptors based on a plurality of retail product tag images corresponding to a retailer category, the plurality of image descriptors including values representative of one or more visual features of the plurality of retail product tag images; generate a probability density function of probability values corresponding to the visual features represented in the plurality of image descriptors of the retailer category; select ones of the probability values based on a sample selection algorithm that, when seeded by a retailer seed value corresponding to the retailer category, generates position values identifying positions in the probability density function of the ones of the probability values to be selected; generate a category signature based on the selected ones of the probability values; and train a convolutional neural network based on a feature descriptor and at least one of the retail product tag images, the feature descriptor including the category signature concatenated to one of the image descriptors, the training of the convolutional neural network to cause the convolutional neural network to classify the at least one of the retail product tag images as one of a plurality of types of product tags

[00100] In Example 9. the subject matter of Example 8 can optionally include that the instructions, when executed, cause the processor to model the probability density function of the probability values with a Gaussian distribution, the Gaussian distribution to approximate a real distribution.

[00101] In Example 10, the subject matter of any one of Examples 8-9 can optionally include that the instructions, when executed, cause the processor to generate the retailer seed value based on a retailer identifier of the retailer category.

[00102] In Example 1 1 , the subject matter of any one of Examples 8-

10 can optionally include that the instructions, when executed, cause the processor to implement a random-number generator to generate the retailer seed as a unique random seed.

[00103] In Example 12, the subject matter of any one of Examples 8-

1 1 can optionally include that the instructions, when executed, cause the processor to generate the category signature as a length representing a number of samples corresponding to ones of the probability values. [00104] In Example 13, the subject matter of any one of Examples 8-

12 can optionally include that the category signature is reproduceable based on the same retailer seed value.

[00105] In Example 14, the subject matter of any one of Examples 8-

13 can optionally include that the instructions, when executed, cause the processor to store the category signature in association with a retailer identifier corresponding to the retailer category, the category signature different from a second retailer category generated by the category signature generator for a second retailer.

[00106] Example 15 is a method that includes generating, by executing an instruction with a processor, a plurality of image descriptors based on a plurality of retail product tag images corresponding to a retailer category, the plurality of image descriptors including values representative of one or more visual features of the plurality of retail product tag images;

generating, by executing an instruction with die processor, a probability density function of probability values corresponding to the visual features represented in the plurality of image descriptors of the retailer category;

selecting, by executing an instruction with the processor, ones of the probability values based on a sample selection algorithm that, when seeded by a retailer seed value corresponding to the retailer category, generates position values identifying positions in the probability density fimction of the ones of the probability values to be selected; generating, by executing an instruction with the processor, a category signature based on the selected ones of the probability values; and training, by executing an instruction with the processor, a convolutional neural network based on a feature descriptor and at least one of the retail product tag images, the feature descriptor including the category signature concatenated to one of the image descriptors, the training of the convolutional neural network to cause the convolutional neural network to classify the at least one of the retail product tag images as one of a plurality of types of product tags

[00107] In Example 16, the subject matter of Example 15 can optionally include modeling the probability density function with a Gaussian distribution to approximate a real distribution.

[00108] In Example 17, the subject matter of any one of the Examples

15-16 can optionally include generating a retailer seed value based on a retailer identifier of the retailer category.

[00109] In Example 18, the subject matter of any one of Examples 15-

17 can optionally include implementing a random-number generator to generate the retailer seed value as a unique random seed.

[00110] In Example 19, the subject matter of any one of Examples 15-

18 can optionally include that the category signature is a length representing a number of samples corresponding to ones of the probability values.

[00111] In Example 20, the subject matter of any one of the Examples 15-19 can optionally include that the category signature is reproduceable based on the same retailer seed value. [00112] In Example 21, the subject matter of any one of the Examples

15-20 can optionally include storing the category signature in association with a retailer identifier corresponding to the retailer category, the category signature different from a second retailer category generated by a category signature generator for a second retailer.

Claims

What is Claimed is:

1. An apparatus comprising:

a feature extractor to generate a plurality of image descriptors based on a plurality of retail product tag images corresponding to a retailer category, the plurality of image descriptors including values representative of one or more visual features of the plurality of retail product tag images;

a probability density function generator to generate a probability density function of probability values corresponding to the visual features represented in the plurality of image descriptors of the retailer category; a sample selector to select ones of the probability values based on a sample selection algorithm that, when seeded by a retailer seed value corresponding to the retailer category, generates position values identifying positions in the probability density fbnction of the ones of the probability values to be selected;

a category signature generator to generate a category signature based on the selected ones of the probability values; and

a processor to train a convolutional neural network based on a feature descriptor and at least one of the retail product tag images, the feature descriptor including the category signature concatenated to one of the image descriptors, the training of the convolutional neural network to cause the convolutional neural network to classify the at least one of the retail product tag images as one of a plurality of types of product tags.

2. The apparatus as defined in claim 1 , wherein the probability density function generator is to model the probability density function of the probability values with a Gaussian distribution, the Gaussian distribution to approximate a real distribution.

3. The apparatus as defined in claim 1, further including a seed generator to generate the retailer seed value based on a retailer identifier of the retailer category.

4. The apparatus as defined in claim 3, wherein the seed generator is implemented by a random-number generator to generate the retailer seed as a unique random seed.

5. The apparatus as defined in claim 1 , wherein the category signature is a length representing a number of samples selected by the sample selector, die samples corresponding to ones of the probability values.

6. The apparatus as defined in claim I , wherein the category signature is reproduceablc based on the same retailer seed value.

7. The apparatus as defined in claim 1 , further including a memory to store the category signature in association with a retailer identifier corresponding to the retailer category, the category signature different from a second retailer category generated by the category signature generator for a second retailer.

8, A non-transitory computer readable storage medium comprising instructions that, when executed, cause a processor to at least: generate a plurality of image descriptors based on a plurality of retail product tag images corresponding to a retailer category, the plurality of image descriptors including values representative of one or more visual features of the plurality of retail product tag images;

generate a probability density function of probability values corresponding to the visual features represented in the plurality of image descriptors of the retailer category;

select ones of the probability values based on a sample selection algorithm that, when seeded by a retailer seed value corresponding to the retailer category, generates position values identifying positions in the probability density function of the ones of the probability values to be selected;

generate a category signature based on the selected ones of the probability values; and

train a convolutional neural network based on a feature descriptor and at least one of the retail product tag images, the feature descriptor including the category signature concatenated to one of the image descriptors, the training of the convolutional neural network to cause the convolutional neural network to classify the at least one of the retail product tag images as one of a plurality of types of product tags.

9. The non-transitory computer readable storage medium as defined in claim 8, wherein the instructions, when executed, cause the processor to model the probability density function of the probability values with a Gaussian distribution, the Gaussian distribution to approximate a real distribution.

10. The non-transitory computer readable storage medium as defined in claim 8, wherein the instructions, when executed, cause the processor to generate the retailer seed value based on a retailer identifier of the retailer category.

1 1. The non-transitory computer readable storage medium as defined in claim 10, wherein the instructions, when executed, cause the processor to implement a random-number generator to generate the retailer seed as a unique random seed.

12. The non-transitory computer readable storage medium as defined in claim 10, wherein the instructions, when executed, cause the processor to generate the category signature as a length representing a number of samples corresponding to ones of the probability values.

13. The non-transitory computer readable storage medium as defined in claim 8, wherein the category signature is reproduceab!e based on the same retailer seed value.

14. The non-transitory computer readable storage medium as defined in claim 8, wherein the instructions, when executed, cause the processor to store the category signature in association with a retailer identifier corresponding to the retailer category, the category signature different from a second retailer category generated by the category signature generator for a second retailer.

15, A method comprising:

generating, by executing an instruction with a processor, a plurality of image descriptors based on a plurality of retail product tag images corresponding to a retailer category, the plurality of image descriptors including values representative of one or more visual features of the plurality of retail product tag images;

generating, by executing an instruction with the processor, a probability density ftinction of probability values corresponding to the visual features represented in the plurality of image descriptors of the retailer category;

selecting, by executing an instruction with the processor, ones of the probability values based on a sample selection algorithm that, when seeded by a retailer seed value corresponding to the retailer category, generates position values identifying positions in the probability density ftinction of the ones of the probability values to be selected;

generating, by executing an instruction with the processor, a category signature based on the selected ones of the probability values; and

training, by executing an instruction with the processor, a convolutional neural network based on a feature descriptor and at least one of the retail product tag images, the feature descriptor including the category signature concatenated to one of the image descriptors, the training of the convolutional neural network to cause the convolutional neural network to classify the at least one of the retail product tag images as one of a plurality of types of product tags.

16. The method as defined in claim 14, further including modeling the probability density function with a Gaussian distribution to approximate a real distribution.

17. The method as defined in claim 14, further including generating a retailer seed value based on a retailer identifier of the retailer category.

18, The method as defined in claim 16, flirther including implementing a random-number generator to generate the retailer seed value as a unique random seed.

19. The method as defined in claim 14, wherein the category signature is a length representing a number of samples corresponding to ones of the probability values.

20. The method as defined in claim 14, wherein the category signature is reproduceable based on the same retailer seed value.

21. The method as defined in claim 14, ftirther including storing the category signature in association with a retailer identifier corresponding to the retailer category, the category signature different from a second retailer category generated by a category signature generator for a second retailer.